Introduction

Do you ever find yourself snapping photos when eating out, almost out of habit?
And then later, when scrolling through your gallery, you think, “Why did I even take this…?” and end up hitting the delete button.

But what if those so-called “useless photos” could actually become useful with the help of AI?
For example, wouldn’t it be handy if AI could estimate calories just from a photo?

So, I decided to put Google’s Gemini to the test.
The experiment: “How accurately can AI estimate calories from photos?”

The Judge: Saizeriya

For this challenge, I chose Saizeriya as the judge.
Why? Because family restaurants like Saizeriya publish detailed allergen and calorie information on their official site, making them perfect for fact-checking.

Reference: 👉 Saizeriya Allergen & Nutrition Information

Now, let’s look at the results.

Young Chicken Diavola Style

Official: 683 kcal

Gemini estimate: 780 kcal


Spicy Chicken Wings

Official: 295 kcal

Gemini estimate: 400–500 kcal


Milano-Style Doria

Official: 550 kcal

Gemini estimate: 600–800 kcal


Squid Ink Sepia Sauce Pasta

Official: 596 kcal

Gemini estimate: 600–700 kcal

As you can see, Gemini consistently overestimated the calories.

Why Did It Fail?

Honestly, I was hoping to wrap this up with a big “See how amazing AI is!”
But reality wasn’t so kind. Here are my thoughts on why it didn’t go well.

1. The “hidden ingredients” problem
For dishes like doria, it’s impossible to tell ingredient proportions from a photo. Gemini seemed to interpret it as just “gratin.”
And to be fair, it does look like one.



2. Ingredient differences
For spicy chicken, I estimated based on 40g per piece.
But maybe the chicken type or cooking style was different.



3. Cooking method guesswork
For the Diavola chicken, I included my hand in the photo to give a sense of scale, but Gemini might have misjudged how it was cooked (like whether excess oil was drained).



Interestingly, pasta dishes turned out relatively better. Likely because:

Ingredients and portions are easier to guess visually

Cooking methods are simpler to identify

Forks or tableware in the photo give size hints

Conclusion

This little experiment started with the idea: “Is there a way to make food photos less useless?”
The calorie estimation attempt, however, turned out to be… a flop.

But here’s what I learned: to make calorie estimates from photos more accurate, you need things like:

Clear size references (hands, chopsticks, or other objects in frame)

Detailed ingredient data for each dish


So, even though this round failed, it actually gave me a clearer picture of what’s missing to make it work.
Next time, I’ll tweak my approach and try again.